Claude Managed Agents Adds a Terminal Session Viewer and Auto Permission Mode

Anthropic has added two practical controls to Claude Managed Agents: a terminal command that attaches to a running session and an auto permission mode that evaluates tool calls against the user's stated intent. The ant beta:sessions connect command can follow a live session, while --web serves a local session viewer. (ClaudeDevs on X; Claude Platform release notes)

The companion auto mode is designed for agents that need to keep moving without turning every tool call into a manual interruption. Anthropic says the server evaluates each agent or MCP tool call based on the intent in user.message and chooses whether to run the call, deny it, or ask for input. (ClaudeDevs on X; Claude Platform release notes)

Key Takeaways

• ant beta:sessions connect attaches a terminal to a running Managed Agents session.
• Adding --web opens a local web session viewer served from the CLI.
• Auto mode evaluates agent and MCP tool calls against the intent carried in user.message.
• The evaluation can run, deny, or pause a tool call for user input.
• Managed Agents events expose the evaluation result, including evaluation and evaluated_permission fields.
• The Console session viewer provides deeper inspection of timelines, transcripts, tools, resources, threads, cost, and token usage for authorized users.

Follow a Session from the Terminal

Managed Agents are useful when a session runs longer than a single interactive request. The new CLI path lets a developer attach to that work after it has started:

ant beta:sessions connect --session-id <session-id>

Anthropic's announcement says the command attaches the terminal to a running session, where a developer can follow the session live, send messages, and allow or deny tool calls waiting for approval. The release notes describe the same command as the way to connect to a Managed Agents session from the terminal. (ClaudeDevs on X; Claude Platform release notes)

That makes the CLI more than a launcher. A developer can start an agent in one workflow, return to it later, and intervene when the session reaches a permission boundary. The command is especially useful for debugging an agent that is running asynchronously or for watching a multi-step task without rebuilding a separate dashboard.

Open the Local Web Viewer

The same command supports a web option:

ant beta:sessions connect --session-id <session-id> --web

Anthropic says --web serves the Claude Console's session viewer locally and opens it there. The viewer is intended for inspecting what an agent did without writing a custom interface. (ClaudeDevs on X; Claude Platform session events and streaming)

The documented Console viewer includes a timeline minimap for navigating activity, a transcript grouped by model request, and an inspector for session details, raw events, tools, resources, and threads. It can show token usage, cost, session budgets, tool failures, and the files written to the session output directory. Anthropic says the viewer is accessible to Developers and Admins. (Claude Platform session events and streaming)

This is an observability improvement with a clear developer payoff: teams can inspect the agent's path, not just its final response. That makes it easier to distinguish a model decision from a tool failure, a permission denial, or a problem in the mounted resources.

Let the Server Evaluate Tool Calls

Auto mode moves part of the permission decision into the Managed Agents server. When an agent emits a tool call, the server evaluates it based on the intent in the user.message event. Anthropic's announcement says the result can be one of three outcomes: run the call, deny it, or ask the user for input. (ClaudeDevs on X; Claude Platform release notes)

The release notes also document the event-level record: agent.tool_use and agent.mcp_tool_use events report an evaluation field alongside evaluated_permission. That gives hosts and debugging tools a way to see not only that a call was made, but how the permission policy evaluated it. (Claude Platform release notes)

Auto mode should not be read as “allow everything.” Its useful distinction is between routine work that matches the request and actions that need a stop. A host can inspect the evaluation, surface a denial, or ask the user to decide when the call falls outside the agent's intended scope. Teams still need to define safe tools, environments, and approval boundaries for the workflows they run.

A Better Fit for Long-Running Agents

The two changes reinforce each other. Auto mode reduces unnecessary pauses for calls that fit the user's intent; the session viewer gives developers a place to inspect the work and handle the calls that do not. Together, they address the two operational costs of long-running agents: waiting for a human at every step and losing visibility into what happened between start and finish.

For a production workflow, a sensible rollout is to begin with the viewer. Inspect the event stream, tool calls, resources, and approval points before enabling an automated evaluation path. Then use auto mode for a narrow class of routine operations, monitor the evaluation and evaluated_permission fields, and keep explicit approval for actions that can change external systems or expose sensitive data. The last part is operational guidance, not a claim that Anthropic has approved a particular deployment policy.

What Developers Can Do Now

A team adopting the update can:

1. Start a Managed Agents session and record its session ID.
2. Attach with ant beta:sessions connect to follow the live run.
3. Add --web when a timeline, transcript, tool inspector, or resource view is more useful than terminal output.
4. Review which calls pause for approval and which are routine.
5. Enable auto mode for a narrowly defined workflow and log the evaluation fields alongside the tool result.

The session viewer also gives teams a concrete debugging artifact. Instead of asking why an agent “got stuck,” a developer can inspect the exact event, permission decision, tool input, and result that caused the pause or failure. (Claude Platform session events and streaming)

Bottom Line

Claude Managed Agents now has a more complete operator loop: attach to a live session from the terminal, open a local web viewer when deeper inspection is needed, and let the server evaluate tool calls against the user's intent. The changes do not eliminate human oversight; they make the handoff between automated execution and approval more observable and more deliberate. (ClaudeDevs on X; ClaudeDevs on X; Claude Platform release notes)

Sources

• ClaudeDevs on X, “Two fresh updates to Claude Managed Agents,” September 10, 2026: https://x.com/ClaudeDevs/status/2098120133549895978
• ClaudeDevs on X, “We've also added auto mode to Claude Managed Agents,” September 10, 2026: https://x.com/ClaudeDevs/status/2098120135718433024
• Anthropic, Claude Platform release notes: https://platform.claude.com/docs/en/release-notes/overview
• Anthropic, “Session event stream — Managed Agents”: https://platform.claude.com/docs/en/managed-agents/events-and-streaming